Microsoft completes MAI-Base-2 at 2.8 trillion parameters — but the efficiency figures are the company's own
Microsoft AI confirmed on Thursday 15 October 2026 that pre-training of MAI-Base-2 is complete and the model is now in post-training. The same day, CEO Mustafa Suleyman told employees that the company's new base model beats the model powering GPT-5.6 Sol — using a quarter of OpenAI's training compute. Every central figure comes from Microsoft itself.
What's new — and who is saying it
At a town hall meeting with employees on Thursday, Suleyman said, according to Business Insider:
"We now have a stronger pre-trained base model than the one currently powering GPT 5.6 Sol. And in fact, we used a quarter of the compute that OpenAI used to train their model."
He added that Microsoft has also "privately and confidentially" compared the model with OpenAI's models, including one nicknamed "Spud," and hinted that the company has special access to information about how OpenAI's models perform. That is an unusually candid disclosure from a meeting that is normally not public, and the comparison rests entirely on Microsoft's own information.
According to Business Insider, Microsoft declined to comment on the matter, and OpenAI did not respond to inquiries. None of the claims have so far been independently verified — neither the efficiency figures nor the comparison with OpenAI.
The model and its training
MAI-Base-2 is a mixture-of-experts model with 2.8 trillion total parameters, of which 80 billion are active per forward pass. According to Microsoft, the model was trained on roughly 50 trillion tokens over 38 days in August and September, and 18 percent of the training data was image material (Investing.com via Yahoo Finance).
The company also says the model outperformed "every open-weight model it was tested against." Which models were tested, and on which evaluation benchmarks, has not been disclosed.
There is an unresolved contradiction in the reporting on the hardware: Investing.com/Yahoo Finance reports 16,000 Nvidia GB200 GPUs, while a flash item from KuCoin/MetaEra, citing AIBase, reports 6,000 GB200 GPUs over the same 38 days. Both figures are company tallies relayed through secondary sources, and the question must for now be considered open. The KuCoin items are relays and should not be counted as independent confirmation.
The efficiency claims
The most concrete figures Microsoft puts forward are:
- According to the company's own evaluations, conducted across four tests, the training method is 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base.
- The same method is 6.3 to 12.7 times more efficient than its predecessor, MAI-Base-1.
Suleyman did not specify which DeepSeek model the comparison refers to, according to KuCoin News, which relays reporting from IT Home and Business Insider.
Microsoft itself attributes the results to three factors: scientific scaling, co-design of model and infrastructure, and data scaling. These explanations are general — they do not describe concrete technical measures or measurable targets that third parties could verify.
A lineage without public weights
What is new in the release strategy is that Microsoft says it will license the model weights to "trusted partners" through Azure AI Foundry, rather than releasing them openly. The company ties this choice to its claim that the model was trained on original data, without knowledge distillation from other systems — a point that is becoming increasingly important for companies that must be able to document provenance and licensing. It is worth noting that none of the available sources explicitly state that the model will not come as open weights; that is an interpretation of the licensing statement, not a confirmed decision.
Neither commercial availability nor pricing has been disclosed. For enterprise customers, the arrangement means the model will most likely be available via Azure, but whether it can be run on customers' own infrastructure — as open-weight models from, among others, DeepSeek allow — remains unclear for now.
The legacy of the MAI program
MAI-Base-2 is the successor to MAI-Base-1, which sits beneath the reasoning model MAI-Thinking-1. The MAI program was launched on 2 June 2026 at Microsoft Build, with seven models, and Microsoft stated at the time that all of them, including MAI-Thinking-1, were built from scratch without the use of third-party models for distillation (Crypto Briefing).
The second-generation base model is thus a continuation of the same story Microsoft has been telling for less than a year: its own base models, its own data, and claims of substantially better compute efficiency than both its predecessor and its competitors.
What remains to be seen
There is so far no primary documentation: none of the reporting sources have seen Microsoft's own announcement or benchmark data. Suleyman's comparison with GPT-5.6 Sol rests on "private and confidential" access to OpenAI technical information, and OpenAI has not commented. The efficiency figures against DeepSeek come from Microsoft's own evaluations.
The next checkpoint will be the model's performance after post-training. Base models are often compared on raw capabilities, but it is the performance of finished, post-trained models that will determine whether Microsoft has really caught up with its competitors — and the independent benchmarks that would reveal that have not yet been run.